Source-linked AI summary

Heterogeneous LoRA for Federated Fine-tuning of On-Device Foundation Models

Yae Jee Cho, Luyang Liu, Zheng Xu, Aldi Fahrezi, Gauri Joshi

arXiv:2401.06432v2cs.LGcs.DC

TL;DR

Federated fine-tuning of on-device foundation models must handle heterogeneous client systems and data while preserving the benefits of parameter-efficient adaptation. The paper proposes HETLORA, which assigns heterogeneous LoRA ranks and combines local rank self-pruning with server-side sparsity-weighted aggregation. HETLORA improves training speed, communication and computation efficiency, and final performance over homogeneous LoRA, while its rank-distribution assumption leaves correlated rank and data distributions for future study.

  • Problem

    Federated fine-tuning of on-device foundation models must accommodate heterogeneous client system capabilities and data distributions, but homogeneous LoRA ranks do not fit this setting well.

  • Method

    HETLORA assigns different LoRA ranks to clients, applies local rank self-pruning, and aggregates heterogeneous modules with sparsity-weighted aggregation.

  • Results

    HETLORA achieves better training speed, communication and computation efficiency, and final performance than homogeneous LoRA.

  • Takeaways & Limitations

    Heterogeneous LoRA provides a practical federated fine-tuning approach for on-device foundation models with varying client resources and data complexity.

  • Takeaways & Limitations

    The method assumes rank distribution across clients is independent of data distribution, although these distributions may be correlated.

Abstract

from arXiv · show

Foundation models (FMs) adapt well to specific domains or tasks with fine-tuning, and federated learning (FL) enables the potential for privacy-preserving fine-tuning of the FMs with on-device local data. For federated fine-tuning of FMs, we consider the FMs with small to medium parameter sizes of single digit billion at maximum, referred to as on-device FMs (ODFMs) that can be deployed on devices for inference but can only be fine-tuned with parameter efficient methods. In our work, we tackle the data and system heterogeneity problem of federated fine-tuning of ODFMs by proposing a novel method using heterogeneous low-rank approximations (LoRAs), namely HetLoRA. First, we show that the naive approach of using homogeneous LoRA ranks across devices face a trade-off between overfitting and slow convergence, and thus propose HetLoRA, which allows heterogeneous ranks across client devices and efficiently aggregates and distributes these heterogeneous LoRA modules. By applying rank self-pruning locally and sparsity-weighted aggregation at the server, HetLoRA combines the advantages of high and low-rank LoRAs, which achieves improved convergence speed and final performance compared to homogeneous LoRA. Furthermore, HetLoRA offers enhanced computation efficiency compared to full fine-tuning, making it suitable for federated fine-tuning across heterogeneous devices.

1 Introduction

The paper studies federated fine-tuning of on-device foundation models under heterogeneous client resources and data, proposing HETLORA to assign different LoRA ranks across clients. HETLORA combines rank self-pruning and sparsity-weighted aggregation to improve efficiency and performance over homogeneous-rank LoRA.

  • Motivation: On-device foundation models require federated fine-tuning because their smaller parameter sizes weaken downstream performance, while private client data cannot always be centralized.ODFMs are defined as models with at most a few billion parameters that fit on limited-capacity clients.
  • Motivation: Existing federated parameter-efficient methods do not adapt PEFT configurations to heterogeneous client system capabilities and data distributions.The paper identifies this as a practical gap for ODFM fine-tuning.
  • HETLORA: HETLORA assigns heterogeneous ranks to clients and uses rank self-pruning plus sparsity-weighted aggregation to accommodate differing capabilities and data complexities.The pretrained weights remain on-device while the server aggregates heterogeneous LoRA modules into a global module.
  • LoRA trade-off: Homogeneous LoRA ranks create a trade-off: larger ranks accelerate training but overfit faster, whereas smaller ranks train more slowly without the same overfitting.LoRA reduces trainable parameters to at most 0.02% of the original ODFM size.
  • Results: HETLORA outperforms homogeneous-rank LoRA in training speed, communication and computation efficiency, and final performance.The evaluation covers PaLM 2 XXS and XS on multi-session chat and Reddit summarization data from clients.

2 Related Work

Prior work combines parameter-efficient fine-tuning methods with federated learning, but limited work adapts LoRA training itself to heterogeneous device capabilities. HETLORA addresses this setting by allowing client-specific LoRA ranks.

  • Parameter-Efficient Fine Tuning: Parameter-efficient fine-tuning freezes most foundation-model parameters while tuning a subset or adding trainable components such as adapters, prompts, prefixes, or low-rank approximations.
  • Federated Fine-Tuning: Recent federated foundation-model work collaboratively trains prompts or adapters, but standard federated settings may not fit smaller on-device models or heterogeneous client capacities.
  • LoRA in Federated Learning: Prior LoRA-for-federated-learning work studies initialization or retains standard LoRA training, without adapting the training process to heterogeneous device capabilities.

3 Federated Fine-Tuning with LoRA

Federated LoRA fine-tuning freezes pretrained ODFM weights while clients train and communicate low-rank modules. HETLORA supports heterogeneous client ranks through truncation, rank self-pruning, and sparsity-weighted aggregation.

  • LoRA Federated Fine-Tuning: LoRA represents the trainable update as ∆W = BA, reducing trainable parameters while keeping the pretrained ODFM weights fixed.The factors have rank r, with B ∈ R^(d×r) and A ∈ R^(r×l).
  • Heterogeneous LoRA: HETLORA assigns each client a rank within [rmin, rmax] to accommodate heterogeneous system capabilities and data complexities.Client ranks may change across communication rounds through self-pruning.
  • Distribution via Truncation: The server distributes global modules by truncating their rank dimension to each selected client’s rank, then clients train locally and return updated modules.This replaces a single fixed rank with client-specific modules while preserving the federated training cycle.
  • Local Training with Rank Self-Pruning: Clients self-prune ranks using a norm-based regularizer that penalizes the last ranks, with γ < 1 controlling pruning intensity.After local training, clients compare updated and initially received module norms and prune the last ranks when the updated norm is smaller.
  • Sparsity-Weighted Aggregation: Sparsity-weighted aggregation reduces the bias of zero-padding toward high-rank clients by weighting reconstructed updates according to their singular-value norms.Naïve zero-padding can overemphasize noisy updates from high-rank clients; HETLORA instead de-emphasizes less informative large-rank updates.

4 Experiments

Experiments evaluate PaLM 2 ODFMs on multi-session chat and Reddit summarization, comparing homogeneous and heterogeneous LoRA with full fine-tuning. HETLORA combines heterogeneous ranks, self-rank pruning, and sparsity-weighted aggregation to improve training speed, final performance, and efficiency relative to homogeneous LoRA.

  • Experimental setup: Experiments use PaLM 2 XXS and XS models on multi-session chat and Reddit summarization tasks, with perplexity and RougeL as evaluation metrics.The LoRA modules are applied to self-attention layers, and clients are sampled per communication round for each task.
  • Homogeneous LoRA: Higher homogeneous LoRA ranks reach better perplexity with fewer communication rounds but overfit faster, whereas rank r = 1 trains more slowly without overfitting.This creates a trade-off between rapid performance achievement and overfitting.
  • Heterogeneous LoRA: Larger heterogeneous minimum ranks speed training but worsen performance, while smaller ranks reduce overfitting, including when rmax = 50.The results attribute this reduced overfitting to the regularizing effect of smaller-rank LoRA modules.
  • HETLORA: Adding self-rank pruning and sparsity-weighted aggregation prevents overfitting and improves training speed and final performance over other baselines.With rmin = 5, the combined method achieves these improvements despite the higher minimum rank.
  • Comparison with baselines: HETLORA with rmin = 5 and rmax = 50 trains faster and performs better than homogeneous ranks r ∈ {5, 50} and reconstruction+SVD, while slightly underperforming full fine-tuning.It also requires significantly fewer trained and communicated parameters than full fine-tuning to reach the targeted Reddit and multi-session chat performance.
  • Ablation: Pruning with γ = 0.99 outperforms no pruning at γ = 1, whereas aggressive pruning at γ = 0.85 harms performance.The reported explanation is that moderate pruning reduces noise from large-rank clients whose data complexity is low.

5 Discussions and Concluding Remarks

HETLORA addresses system and data heterogeneity in federated fine-tuning of on-device foundation models. It improves training speed, efficiency, and final performance over homogeneous LoRA, while leaving rank assignment and theoretical analysis for future work.

  • HETLORA caters to device system and data heterogeneity in federated fine-tuning of on-device foundation models.
  • HETLORA achieves better training speed, communication/computation efficiency, and final performance than homogeneous LoRA.
  • Increasing ranks does not always improve homogeneous LoRA, consistent with prior findings.
  • Effective rank assignment across clients for better convergence and performance remains an open question.
  • Theoretical convergence and generalization of heterogeneous LoRA remain important next steps.

6 Limitations

The method assumes that rank distribution across clients is independent of data distribution, although these distributions may be correlated in practice.

  • HETLORA assumes rank distribution across clients is independent of data distribution.
  • Future work should examine correlated rank and data distributions to understand their implications for heterogeneous LoRA.
Loading 2401.06432v2…